We introduce command-line assistants as LLM-based systems that can do real work through tools, permissions, and sandboxed execution environments. We explain how these assistants differ from plain chatbots, opening new possibilities but also introducing challenges and risks.

The main focus is on existing and possible applications in our department.